Trajectory Prediction using Conditional Generative Adversarial Network

Optimization based planners (OBP) use a linear initialization as a prior of their optimizations which fails to use already acquired knowledge. Most of the time the linear initialization will collide with obstacles which will be the most difficult part of the OBP to optimize. We propose a method to perform trajectory prediction that leverages motion dataset by using a conditional generative adversarial network. Unlike previous methods, our proposed method does not require the dataset during execution time but instead generate new trajectories. We demonstrate the validity of our method on simulation. Our method decreases by 20% the number of colliding trajectories predicted compared to the linear initialization while being very fast.